A two-stage causality method for time series prediction based on feature selection and momentary conditional independence

Since the actual time series contain a lot of variables and the relations among them are complex. Hence, it is difficult to accurately judge cause and effect by conventional causality methods. Aiming at the problem, a two-stage causal network learning method, the feature selection stage and the conditional independence test stage, is proposed to reveal the causalities between variables and construct an accurate prediction model. In the first stage, there are two steps to perform. Firstly, a feature selection method is utilized to reduce data dimensionality by removing irrelevant and redundant variables. These variables are not only increase computational complexity, but also cover up part of the effective information, which may result in insufficient accuracy of the constructed model. Then, a global redundancy minimization (GRM) scheme is used to further refine the result of the previous step from a global perspective. In the second stage, a momentary conditional independence (MCI) test is performed to test the causalities between variables, which can accurately detect the causal network structure. Finally, an accuracy causal network and subsequent prediction model can be established based on the output of the two-stage model. In this simulations, two benchmark datasets, a coupled Lorenz system and two actual datasets are used to verify the effectiveness of the proposed method. The results show that the proposed method can effectively analyze the causalities between variables and construct an accuracy prediction model.